From capture to decision
The rest of this area describes surfaces. This page describes the arc they form, because the individual capabilities make more sense once you can see what they are collectively for.
One sentence carries it:
An observation that arrives with its own proof, is understood only where that is lawful and possible, and can be refused at any point, is a governed observation system rather than an ingest pipeline.
Three movements follow from that, and each has a part that is genuinely unusual and a part that is honestly missing.
1. Capture: findings, not bytes
The instinct with sensors and cameras is to move the data to the platform. This one does close to the opposite, and it is the design decision the rest depends on.
A device ships signed digests, not pixels. A capture bundle carries the hash of the sample, the previous hash, the sensor, the sequence number, the capture timestamp and a signature. It does not carry the sample bytes. The recording stays on the device, and what travels is the evidence that it happened, exactly when it says, unaltered.
That has consequences worth stating plainly:
| Property | What it means |
|---|---|
| Two independent layers | A hash chain gives order and integrity; a per-capture signature gives origin. Breaking one does not forge the other |
| Verified at the cell, per capture | Each signature is re-checked against the device's own certificate. An unverified capture is never persisted |
| The chain is re-walked on arrival | After writing, the cell reads the session's chain back and re-derives it. A chain that no longer re-derives raises and refuses rather than quietly holding a reordered history |
| Refused, not re-shipped | A bundle that has been spooled through an outage and no longer verifies is kept as evidence, not retried |
| Metered | Capture ingestion is an accounted operation, not free background traffic |
Bandwidth is the obvious benefit and the least interesting one. The real result is that provenance is cheap enough to be universal. You can afford to sign every capture when a capture is a few hundred bytes.
No server-side analysis of the original media, because the original never arrives. Analysis happens on the device or not at all. That is the trade, and for a sovereign or bandwidth-constrained deployment it is usually the right one.
2. Understanding: where it runs is the design
There is no single "the platform understands media" answer, and a reader who wants one will be misled. The honest shape is that understanding is split by where it can run.
On the device
An Edge Agent carries real inference, on CPU, with no accelerator assumed:
- Object detection, optical character recognition, speech to text, and a local language model, all compiled into the shipped binary.
- Each one fails closed on an unprovisioned model. The engine ships; the model arrives as a signed bundle applied to the device's registry. With no model, the tool refuses rather than returning a plausible answer.
- Results are recorded as device-signed perception records, best-effort: if the device is unenrolled or the local store will not open, the result still returns, unsigned.
In the cell
- Optical character recognition is wired end to end, profile-gated, with a signed observation written back and the operation metered.
- Documents are read across 38 formats and become a retrievable corpus, then entities against your own governed ontology, promoted into a bitemporal, signed, evidence-linked knowledge graph.
- Video is catalogued, not decoded. There is no decoder on the platform.
What is honestly not there
Naming these is not a disclaimer, it is what makes the rest usable for scoping:
| Gap | State |
|---|---|
| Cell-side audio transcription | Engine, signed sink and route each exist; nothing composes them |
| Video decoding | No decoder. Container, size and hash only |
| Cameras, microphones, RTSP, ONVIF | Implemented and compiled in; no production construction site |
| Automatic identity promotion | Surveillance subjects and candidate identities are read-only in production; nothing writes them |
| Open-vocabulary entity extraction | Built with cost gates and circuit breakers; its endpoint refuses |
3. Refusal is part of the product
This is the movement most likely to be mistaken for a limitation, and it is the one worth leading a conversation with.
The platform will tell you what it may not do, and why. A deployment can ask for a per-capability readiness report, and each capability comes back as one of six states: ready, engine missing, needs a model, gating required, off with no policy, or prohibited, each with a reason and a legal basis.
The ordering is the point:
Compliance dominates technical presence. A prohibited capability is reported as prohibited whether or not a model is installed. You cannot make a prohibited capability ready by installing a model.
Three consequences of that, all shipped:
- Emotion and sentiment inference is compiled out of shipped builds, cited against the EU AI Act's prohibition on emotion inference in the workplace and education. Not disabled by a flag someone can flip. Absent from the binary.
- Biometric modalities fail closed to off when no jurisdiction is configured. The absence of a policy is not read as permission.
- Nine media skills refuse rather than degrade. Face recognition, gait, speaker identification, licence-plate reading, weapon detection and others are in the catalog and raise when invoked, rather than quietly falling through to a text extractor and returning something that looks like an answer.
That last one is a deliberate anti-fabrication decision, and it is the difference between a system that says "I cannot do that" and one that hands back a confident guess.
How the arc composes
Figure: the arc. Capture produces evidence rather than media (neutral); verification is a gate that refuses rather than warns (amber); understanding runs on the device or in the cell and converges on one provenance record (green); and the readiness and policy layer decides whether a capability may run at all, answering with a reason and a legal basis when it may not. Ceiling: this is the shape of the wired path; the gaps named above sit outside it.
Table equivalent
| Stage | Produces | Refuses when |
|---|---|---|
| Capture | A signed, chained digest | The signature does not verify |
| Verify | A persisted provenance observation | The chain does not re-derive |
| Understand (device) | A device-signed perception record | No model is provisioned |
| Understand (cell) | A signed observation, entities, a corpus | The gate is closed, or the engine is absent |
| Readiness and policy | Permission to run at all | Prohibited, gated, or unpoliced |
| Decision | An action with an audit record | Any gate above said no |
Related
- Sensor intake, the capture plane in depth.
- Surveillance, the observed-to-canonical identity ladder and its human gate.
- Unstructured data and media, the document and media path in the data plane.
- Supported file types, every format read and identified.
- Core and edge, the two runtimes this arc spans.
- Classification, the clearance that bounds who sees an observation.